optimal use of land surface temperature data to detect ... · optimal use of land surface...

Optimal use of land surface temperature data to detect changesin tropical forest cover

Thijs T. van Leeuwen,1,2 Andrew J. Frank,3 Yufang Jin,1 Padhraic Smyth,3

Michael L. Goulden,1 Guido R. van der Werf,2 and James T. Randerson1

Received 26 July 2010; revised 31 December 2010; accepted 18 January 2011; published 15 April 2011.

[1] Rapid and accurate assessment of global forest cover change is needed to focusconservation efforts and to better understand how deforestation is contributing to thebuildup of atmospheric CO2. Here we examined different ways to use land surfacetemperature (LST) to detect changes in tropical forest cover. In our analysis we usedmonthly 0.05° × 0.05° Terra Moderate Resolution Imaging Spectroradiometer (MODIS)observations of LST and Program for the Estimation of Deforestation in the BrazilianAmazon (PRODES) estimates of forest cover change. We also compared MODIS LSTobservations with an independent estimate of forest cover loss derived from MODISand Landsat observations. Our study domain of approximately 10° × 10° included theBrazilian state of Mato Grosso. For optimal use of LST data to detect changes in tropicalforest cover in our study area, we found that using data sampled during the end of the dryseason (∼1–2 months after minimum monthly precipitation) had the greatest predictiveskill. During this part of the year, precipitation was low, surface humidity was at aminimum, and the difference between day and night LST was the largest. We used thisinformation to develop a simple temporal sampling algorithm appropriate for use inpantropical deforestation classifiers. Combined with the normalized difference vegetationindex, a logistic regression model using day‐night LST did moderately well at predictingforest cover change. Annual changes in day‐night LST decreased during 2006–2009relative to 2001–2005 in many regions within the Amazon, providing independentconfirmation of lower deforestation levels during the latter part of this decade as reportedby PRODES.

Citation: van Leeuwen, T. T., A. J. Frank, Y. Jin, P. Smyth, M. L. Goulden, G. R. van der Werf, and J. T. Randerson (2011),Optimal use of land surface temperature data to detect changes in tropical forest cover, J. Geophys. Res., 116, G02002,doi:10.1029/2010JG001488.

1. Introduction

[2] The clearing of humid tropical forest has importantconsequences for ecosystem function [DeFries et al., 2004;Foley et al., 2005], regulation of regional [Nobre et al.,1991] and global climate [Shukla et al., 1990; Bala et al.,2007], water resources [Hamilton and King, 1983], bio-geochemical cycles [Houghton, 1991], biological diversity[Sala et al., 2000], and the maintenance of soil fertility[Andreux and Cerri, 1989; Davidson et al., 2007]. Sub-stantial improvements in the quantification of tropical forestcover losses have beenmade during the last decade [Houghton

et al., 2000; Achard et al., 2002; Instituto Nacional dePesquisas Espaciais (INPE), 2002; DeFries et al., 2002;Food and Agriculture Organization, 2006; Hansen et al.,2008], and with the likelihood of new climate agreementsincluding Reducing Emissions from Deforestation and Deg-radation (REDD) [United Nations Framework Conventionfor Climate Change, 2005] there will be an even greaterneed for accurate forest monitoring methods, particularly forapproaches that enable monitoring in near real time [e.g.,Morton et al., 2005]. This information is required to betterunderstand land use change effects on the global carbon cycle[Canadell et al., 2007], and to gauge the success of mitigationactivities involving forests [Nabuurs et al., 2007]. A primarychallenge in estimating contemporary rates of forest coverloss is to efficiently extract information from different satel-lite products at multiple spatial and temporal resolutions andthen to combine this information in an effective way withground‐based observations [Achard et al., 2007; Goetz et al.,2009; DeFries et al., 2007].[3] To map changes in tropical forest cover, often a first

step is to classify land cover types, including different types

1Department of Earth System Science, University of California, Irvine,California, USA.

2Department of Hydrology and Geo‐environmental Sciences, VUUniversity Amsterdam, Amsterdam, Netherlands.

3Department of Computer Science, University of California, Irvine,California, USA.

Copyright 2011 by the American Geophysical Union.0148‐0227/11/2010JG001488

JOURNAL OF GEOPHYSICAL RESEARCH, VOL. 116, G02002, doi:10.1029/2010JG001488, 2011

G02002 1 of 16

http://dx.doi.org/10.1029/2010JG001488

of forest, savanna, and agriculture. Surface reflectances inthe visible and near infrared, and vegetation indices (VI),which are linear or nonlinear combinations of the reflec-tance in two or more bands, are frequently used as metricsin land cover classification algorithms [e.g., DeFries andTownshend, 1994; Los et al., 1994]. Annual time series ofthe normalized difference vegetation index (NDVI), basedon AVHRR data, were used by DeFries and Townshend[1994] to develop global land cover maps with a 1° spa-tial resolution. As global AVHHR data became available athigher resolution [Eidenshink and Faundeen, 1994], landcover maps with a finer spatial resolution (8 km and 1 km)were created [DeFries et al., 2000; Hansen and DeFries,2004]. The accuracy of land cover change estimates hasbeen further enhanced by the development and applicationof more sophisticated classification algorithms, includingdecision trees [Hansen et al., 2000; Friedl and Brodley,1997], neural networks [Gopal and Woodcock, 1996], andmixture models [DeFries et al., 1995; Anderson et al.,2005]. In several different studies fractional cover imagesderived from mixture models have been used successfullyfor monitoring deforestation [Shimabukuro et al., 1998] andland cover change [Carreiras et al., 2002].[4] Although surface reflectance observations and VIs

are widely used in assessments of forest cover, additionalinformation can be extracted from thermal infrared bands.Land surface temperature (LST) during the day, derivedfrom thermal infrared bands, has been shown to be closelyrelated with the density of the canopy across differentvegetation types [Price, 1984; Smith and Choudhury, 1991;Nemani and Running, 1997] and therefore helpful in classi-fying land cover types [Lambin and Ehrlich, 1995; Nemaniand Running, 1997; Roy et al., 1997]. Borak et al. [2000]confirmed the importance of LST data as a complementarysource of information to the NDVI data. In more recentstudies the daytime land LST product of the MODIS Terrasatellite has been included in algorithms for detecting forestcover change in the humid tropics [e.g., Hansen et al., 2008].The information content of LST, however, has not beensystematically evaluated [Mildrexler et al., 2007]. Keyuncertainties remain with respect to how the effectivenessof LST information varies seasonally and how it compareswith information derived from visible and near infraredwavelengths.[5] Changes in land cover influence LST by means of

several different pathways [e.g., Carlson and Gillies, 1994;Friedl, 2002; Pongratz et al., 2006]. LST is regulated by theamount of shortwave radiation absorbed by the surface (i.e.,surface albedo), surface conductance, the amount of wateravailable for evaporative cooling, wind speed, and surfaceroughness which regulates the strength of both sensibleand latent heat fluxes. During the dry season, daytime LSThas been shown to be linearly related to the density of thecanopy across different vegetation types [Nemani et al.,1993]. Areas of tree cover, which often have deeper rootsand thus may access greater water resources, tend to havehigher rates of evapotranspiration (and thus lower LSTs)during the dry season than grasses and other herbaceouscover that may senesce. Increases in LST in deforested areasmay be further amplified by reductions in roughness lengththat reduce the dissipation of energy by means of eithersensible or latent heat fluxes. Different model experiments

in the state of Mato Grosso, Brazil, conducted by Pongratzet al. [2006] confirm that day LST increases in areas withlower forest cover, particularly during the dry season forpastures, compared with dense forest.[6] Night LSTs are also sensitive to forest cover. Noc-

turnal drainage of air from upper canopy layers and poolingof cold air at the forest floor keeps upper levels of thecanopy relatively warm [Goulden et al., 2006]. This processof nocturnal cold air drainage and pooling cannot take placein short stature vegetation; net loss of thermal radiationcools the land surface during the night, but the colder airthat is created from interaction with the canopy remains atthe surface and in contact with the canopy elements thatare emitting thermal radiation. Satellite sensors only mea-sure the temperature of the top of forest canopies, so intactforest shows a higher night LST than comparable areas withshort‐stature vegetation [Goulden et al., 2006]. Consideringboth day and night LSTs, differences in the diurnal tem-perature range are expected to be smaller for forests andlarger for grasslands and shrublands, given the biophysicalprocesses described above [Collatz et al., 2000]. WithinMato Grosso, Pongratz et al. [2006] found, for example,that the diurnal temperature range increases by over 7°Cwhen forests are converted to bare ground.[7] Here we explored the use of LST data from the

MODIS land surface climate modeling grid (CMG) productfor quantifying forest cover and its change, and comparedour findings to PRODES estimates of deforestation lossesin the state of Mato Grosso, Brazil. We assessed how theinformation content of LST as a predictor of forest coverchanged seasonally and also as a function of day, night,and day‐night LST differences. Based on this analysis, wedeveloped a metric to monitor forest cover change in thehumid tropics. Several experiments were performed to fur-ther test the ability of day LST, night LST, and day‐nightLST differences during different seasonal periods for forestcover mapping by ingesting them one by one in a logisticregression model. We found that effective use of the LSTdata required identifying the optimal seasonal period withineach region when water vapor in the atmosphere was minimaland drought stress was greatest. Our study demonstrated theusefulness of MODIS LST data at a coarse spatial resolutionfor the rapid identification of deforestation hot spots and forassessing regional trends.

2. Methods

2.1. Study Area

[8] Our study area encompassed the state of Mato Grosso,Brazil (7.0°–19.0°S, 62.0°–50.0°W) (Figure 1a). Evergreenbroadleaf forest was the dominant plant functional type (pft)in the northern part of the state. Woodland savanna (cerrado),savanna, and agriculture pfts were more abundant in thesouthern and eastern part of the domain (Figure 1b). One ofthe largest contiguous tracts of forest in Mato Grosso occurswithin the Xingu Indigenous Reserve, located in the north-eastern part of the state (red line in Figure 1b). One of theworld’s largest wetlands, the Pantanal, is located in the southcentral part of the domain (gray line in Figure 1b).[9] Located within the “arc of deforestation,” the state of

Mato Grosso has experienced rapid changes in land coverbecause of agricultural expansion and intensification [Morton

VAN LEEUWEN ET AL.: USING LST TO DETECT FOREST COVER CHANGE G02002G02002

2 of 16

et al., 2006]. The two most dominant forms of land coverchange in Mato Grosso during 2001–2004 were forest con-version to pasture and forest conversion to cropland,accounting for approximately 65% and 20%, respectively,of all observed land cover transitions [Morton et al., 2006].Areas previously cleared for cattle ranching were alsoobserved to undergo subsequent conversion to soybeanagriculture (agricultural intensification), and to a lesserextent, to corn, cotton and sugarcane agriculture. Remaining

forest and woodland savanna in central, eastern, and southernMato Grosso may be vulnerable to further change as a resultof continued expansion of mechanized agriculture [Jasinskiet al., 2005]. The large mechanized clearings that are usedfor the creation of pastures and croplands in the study areaare more easily detected by the coarse resolution productspresented in this study.

2.2. Data

2.2.1. MODIS LST and NDVI[10] We obtained LST and NDVI products from the

Moderate Resolution Imaging Spectroradiometer (MODIS[Salomonson et al., 1989]), on board NASA’s Terra satellite(nominal 10.30 A.M. descending and 10:30 P.M. ascendingequatorial crossing times). We used data from MODIS Terra(instead of Aqua) because our initial focus was on the 2000–2005 period; this allowed us to make direct comparisonswith theHansen et al. [2008] deforestation estimates describedbelow.[11] For LST we used the collection 5 monthly level‐3

MOD11C3 product described by Z. Wan (MODIS LandSurface Temperature Products Users’ Guide, http://www.crseo.ucsb.edu/modis/LstUsrGuide/MODIS_LST_products_Users_guide_march06.pdf, 2006). This product providesmonthly composited and averaged temperature and emissiv-ity values at a 0.05° geographic climate modeling grid (CMG),as well as the averaged observation times and viewing zenithangles for daytime and nighttime LSTs. The product is derivedfrom the daily MOD11C1 product that has a 0.05° × 0.05°spatial resolution, which is, in turn, based on a reprojection andaverage of the MOD11B1 product.[12] Validation of the MOD11C3 product has occurred

through a series of field campaigns conducted in 2000–2006and over more locations and time periods through radiance‐based validation studies [Wan et al., 2002]. Clouds preventsurface thermal infrared radiation from reaching the satellite,and therefore the MODIS LST data are only available forclear‐sky conditions. We created a spatial subset over the12° × 12° region (240 × 240 grid cells) of our study domainusing all available data from March 2000 through December2005. We also compared changes in LST during two sep-arate periods: 2001–2005 and 2006–2009 to test whetherrecent decreases in deforestation reported by PRODES areconsistent with the LST observations.[13] For NDVI we used the monthly Level‐3 MOD13C2

product that was reprojected on a 0.05° geographic CMG[Huete et al., 2002]. The data are spatial and temporalcomposites of the cloud‐free gridded MODIS NDVI datathat are provided every 16 days at a 1 km resolution(MOD13A2). The historical MODIS NDVI climatologyrecord is used to fill those pixels without any cloud‐freeobservations to achieve cloud‐free global coverage.[14] The MODIS NDVI algorithm operates on a per‐pixel

basis and ingests multiple level‐2 daily atmosphericallycorrected surface reflectances to generate a compositedNDVI (A. Huete et al., MODIS Vegetation Index (MOD13):Algorithm Theoretical Basis Document, http://modis.gsfc.nasa.gov/data/atbd/atbd_mod13.pdf, 1999). It first filtersthe data based on quality, cloud and viewing geometry.Observations that are contaminated by residual atmosphericeffect and extreme off‐nadir views are considered as lowquality data. Only the higher quality and cloud‐free filtered

Figure 1. (a) The spatial domain of our analysis includedthe state of Mato Grosso, Brazil. (b) Evergreen broadleafforests were the dominant land cover type in the northernpart of the domain. Toward the south and east, woodlandsavannas, savannas, and agriculture land cover classes wereincreasingly abundant. The perimeters of the Pantanal andXingu Indigenous Reserve are indicated by the gray andred lines, respectively. These observations are from theMODIS Land Cover Product MOD12C1 [Friedl et al.,2002] for the year 2004 and with a spatial resolution of0.05° × 0.05°.


3 of 16

data are retained for temporal compositing. The two highestNDVI values during each 16 days are then compared andthe one closest to nadir view is selected. This constrainedview‐angle composite method reduces the spatial andtemporal inconsistency due to angular variations encoun-tered in the traditional maximum value composite (MVC)[Holben, 1986].2.2.2. PRODES Forest Cover[15] The Program for the Estimation of Deforestation in

the Brazilian Amazon (PRODES) is one of the largest forestmonitoring projects in the world [INPE, 2002]. The PRODESdeforestation product is generated from the analysis of high‐resolution Landsat Thematic Mapper (TM) images by theBrazilian Instituto Nacional de Pesquisas Espaciais (NationalInstitute for Space Research, INPE). Areas of deforestation ofmore than 6.5 ha can be detected. The first digital mappedversion was created in 1997, and since 2000 the product hasbeen produced annually with a time delay of approximately5 months after the end of each calendar year. The PRODESdata set covers the Brazilian Amazon.[16] PRODES data that we used in this study were

downloaded at 90 m resolution in GeoTiff format from theINPE website (http://www.obt.inpe.br/prodes/). To reprojectthe PRODES observations onto the same geographic grid asthe MODIS observations, we first converted the PRODESGeoTiff image to a vector (polygon) format using ArcGISsoftware. We then reprojected these polygons from DatumSAD69 to WGS84 using the SAD_1969_To_WGS_1984_1geographic transformation method that yielded a registrationerror that was less than 10m. The final reprojected andregridded PRODES forest cover product for 2000 is shownin Figure 2a. Forest cover loss between the years 2000 and2005 is shown in Figure 2b. With the base map of forestcover in 2000 and the increments of annual deforestation,we also calculated the forest cover for each year during2000–2005.2.2.3. Hansen et al. [2008] Forest Cover Loss[17] Besides the PRODES data set, we included a com-

parison between LST derived forest cover and forest coverchange with the forest cover loss data set of Hansen et al.[2008]. This data set represents gross forest cover loss forthe humid tropical forest biome between the year 2000 and2005. MODIS and Landsat data were combined to estimateareas where forest clearing occurred. Moderate resolutionMODIS data were used to identify areas of likely forestcover loss and to stratify the humid tropics into regions oflow, medium, and high probability of forest clearing. Varioussamples sized 18.5 km × 18.5 km were taken within theseareas, and interpreted for forest cover and forest clearing byusing high spatial resolution Landsat imagery from 2000 and2005. A more detailed description of the data set is given byHansen et al. [2008].[18] The humid tropical forest cover loss data that we used

in this study were downloaded in a sinusoidal GeoTiffformat at 18.5 km × 18.5 km spatial resolution from theGlobal Forest Monitoring website of South Dakota StateUniversity (http://globalmonitoring.sdstate.edu). To reprojecttheHansen et al. [2008] data set onto the same geographic gridas the MODIS observations, we followed the procedure aswith converting PRODES data, as described in section 2.2.2.Hansen et al. [2008] forest cover loss for our study areabetween the years 2000 and 2005 is shown in Figure 2c.

Figure 2. (a) Forest cover in Mato Grosso for the year2000 from the Program for the Estimation of Deforestationin the Brazilian Amazon (PRODES [INPE, 2002]). Theoriginal PRODES data were resampled to a geographic0.05° × 0.05° grid and are shown here as a fraction of totalforest cover between 0 and 1. White pixels correspond to thearea outside the PRODES domain. (b) Forest cover lossbetween the years 2000 and 2005 from PRODES. (c) Forestcover loss between the years 2000 and 2005 in Mato Grossofrom Hansen et al. [2008]. White pixels correspond to thearea outside the Hansen et al. humid tropical forest coverchange domain.


4 of 16

2.2.4. Precipitation and Relative Humidity[19] In this study we used the Tropical Rainfall Measuring

Mission (TRMM) 3B43 product to estimate precipitationwithin our study domain. The TRMM 3B43 algorithm usesfour different independent data streams to produce the bestestimate precipitation rate (mm/h) and root‐mean‐square(RMS) precipitation error estimates [Kummerow et al.,1998]. These data have a spatial resolution of 0.25° × 0.25°and a monthly time step. We resampled the TRMM productduring 2000–2005 at a spatial resolution of 0.05° × 0.05° tooverlay on the LST and NDVI CMG grid.[20] To estimate the relative humidity level at the surface

(1000 hPa) in our study area, we used data from the AIRSinstrument, on board of NASA’s Aqua satellite. In our studywe used the level 3 V5 monthly AIRX3STM product, with aspatial resolution of 1° × 1°. The product contains standardrelative humidity (in %) for different pressure levels that isderived from observations from the Atmospheric InfraredSounder (AIRS) and the Advanced Microwave SoundingUnit (AMSU) [Aumann et al., 2003]. For this study we usedthe relative humidity data at 1000 hPa from the ascendingorbit. These observations represent relative humidity duringthe day at a ∼1:30 P.M. overpass time.[21] The PRODES forest cover map for 2000, reprojected

to the CMG (section 2.2.2), was used to separate forest andnonforest areas using a threshold of 0.5 fractional cover.Due to the coarse resolution of the AIRS product (1° × 1°)compared to the PRODES forest cover, only AIRS pixelsthat fell in the forest domain (forest cover greater than 0.5)were used to estimate mean surface humidity in the forestpart of the domain. Relative humidity of the nonforest wascalculated by using pixels that fell completely in the non-forest domain (forest cover less than or equal to 0.5). TheAIRS product was available from August 2002 through thepresent, and for our analysis we constructed a mean annualcycle for years 2003–2005.

2.3. Logistic Regression

[22] We used a logistic regression model [Hosmer andLemeshow, 2004] to relate MODIS LST and NDVI observa-tions to per‐pixel proportional forest cover. More precisely,we defined the model as

lnpit

1� pit

� �¼ �TXit ð1Þ

where pit is the fraction of forest cover for pixel i in year taccording to PRODES (see section 2.2.2), Xit is a vector ofMODIS observations for pixel i in year t, and b is a vector ofmodel parameters (which are estimated from a set of trainingdata). In the most general case we consider, the vectorsXit andb each have three components: the first corresponding to aconstant intercept term, the second to a NDVI measurement,and the third to a LST measurement. The index i ranges overspatial pixels and the index t ranges over years. Each pixel iand year t provided one training data point, i.e., a Xit and pitpair. To define a NDVI or LST Xit, we computed the meanvalues of the corresponding product either over a 12 monthperiod from February to January, or for a locally defined3 month period during the dry season or wet season (seesection 3.3 for details). Given an estimated parameter vectorb

the model made predictions about proportional forest cover(e.g., at a future time t) according to:

pit ¼ 1

1þ e��TXitð2Þ

To estimate the parameters we used the traditional iterativelyreweighted least squares (IRLS) algorithm for logisticregression [Nelder and Wedderburn, 1972]. This algorithmcorresponds to a Newton‐Raphson optimization of the like-lihood function. The vector b is computed such that thelikelihood of the training data is maximized and then used inequation (2) to make predictions of forest cover in the testdata. Since we were interested in the forest cover change (i.e.,deforestation and forest regrowth), we made predictions bothat the beginning and end of a period of interest and estimatedchanges in forest cover as the difference between the twopredicted forest cover maps. We also evaluated an alternativeapproachwhere a logisticmodel was trained using differencesin MODIS signals to predict differences in forest proportionsmore directly, but we found that this produced no significantdifference in prediction accuracy compared to computing thedifference of two predictions at different times.[23] By evaluating the model’s performance given several

different sets of input features, we assessed the value of dayLST, night LST, and day‐night LST differences as metricsfor detecting forest cover change. In each of our modelexperiments, we trained the model using data from years2000 through 2004. We made predictions both for year 2000and year 2005 to estimate the difference in forest coverbetween the two years. To evaluate model performance, wecompared the estimated difference with the deforestationrates reported by PRODES. Note that PRODES does notinclude forest regrowth (an increase in forest cover between2000 and 2005) whereas in the model these transitions wereallowed.[24] We ran a series of experiments varying the input fea-

ture data source (LST alone, NDVI alone, or both together),the time of day at which the LST data were collected (day,night, or day‐night difference), and the sampling period forconstructing the feature space (dry season, wet season, or12 month mean). In our first experiment we restricted our-selves to dry season LST data and evaluated model perfor-mance for each of day, night, and day‐night differencemeasurements. In our second experiment we used day‐nightdifference LST data and examined the effect varying thesampling period. Next we trained the model using only theNDVI data, again for all three of the different samplingperiods. Finally, we evaluated the model using day‐nightdifference LST data together with NDVI, once again com-paring all three sampling periods.[25] For each experiment we computed several perfor-

mance metrics to help compare the relative quality of thepredictions. First, we considered the linear correlationbetween the predicted and actual forest cover change valuesover the entire region of study. We then considered the ratioof the predicted area versus actual change values: a valueof 1 would be ideal, and values above or below 1 indicatethat the model is either too sensitive or unresponsive,respectively, to changes in the input features. We also esti-mated the linear correlation between predicted and actualchange values over the active deforestation domain: the


5 of 16

subset of pixels for which PRODES indicates there wasforest cover loss between 2000 and 2005. Finally, we com-puted the ratio of predicted to actual deforested area for theactive deforestation domain.

3. Results

3.1. Spatial and Temporal Patterns of LST

[26] Mean daytime LST within the domain was substan-tially lower in forest areas than in nonforest areas through-out the year (Figure 3a). The monthly daytime temperaturewas at a maximum in August for forest areas (300.7K ± 1.2K(1s)) and in September for nonforest areas (306.2K ± 2.2K).The largest forest‐nonforest day LST differences occurrednear the end of the dry season in August (4.4K), September(6.0K), and October (5.5K) (Figure 3a). The spatial distri-bution of LST was inversely related to forest area, as shownin Figure 4a. In August, the highest day LSTs, averagedduring 2000–2005, occurred in nonforest areas in the south-eastern part of the study area. The lowest LSTs occurred inforests in the northwest (Figure 4a).[27] During night, forest areas in the north showed higher

temperatures than those in the southern part of the studyarea, except for the wetlands of the Pantanal (Figure 4c). Theannual cycle indicated that nonforest areas had more coolingat night than forest areas from May to September, with thegreatest difference observed during July. The nighttimeLST minimum within the domain occurred during the samemonths as the precipitation minimum, with a mean of 292.8 ±0.5K for forest areas and 290.2K ± 1.4K for nonforest areas inJuly (Figure 3a). From October to April, when precipitationand surface humidity levels were high, nighttime tempera-tures were almost the same (around 294K) for forest andnonforest areas. This pattern may have been caused bydecreased losses of longwave radiation (and thus decreasedsurface cooling) from high levels of column water vapor.In addition to the trapping of outgoing longwave radiation,higher soil moistures during the wet season may increase theheat capacity of the soils, further reducing the day‐nightsurface temperature differences observed in grasslands andcroplands [e.g., Smith and Choudhury, 1991; Bruno et al.,2006]. The seasonal amplitude of LST was smaller in forestareas than that in nonforest areas for both day and nightperiods.[28] Day‐night LST differences were the largest toward

the end of the dry season, with maxima during August of7.6K ± 1.4K for forest areas and 14.2K ± 2.8K for nonforestareas (Figure 3b). A spatial map of the maximum day‐nightLST difference for August is shown in Figure 4e. In thesoutheastern part of our study area the largest day‐nightdifferences were observed. Smaller differences occurred inthe northern forests and in the Pantanal region.[29] Precipitation within the domain was at a minimum

during June and July. In forest areas the minimum precipi-tation was in July with a mean of 18.9 mm/month (Figure 3c).The precipitation minimum in nonforest areas occurred amonth earlier and had a lower mean (12.4 mm/month).Annually, forest areas received on average 24% more pre-cipitation than nonforest areas. The 1–2 month time delayobserved between minimum monthly precipitation and min-imum monthly humidity may partly reflect increasing waterstress throughout the dry season that caused decreases in

Figure 3. Surface temperatures, precipitation, and surfacehumidity for forest and nonforest parts of the study area.(a) Mean monthly day and night land surface temperatures(K) during 2000–2005 from the MODIS MOD3C11 product(Z. Wan, MODIS Land Surface Temperature Products Users’Guide, http://www.crseo.ucsb.edu/modis/LstUsrGuide/MODIS_LST_products_Users_guide_march06.pdf, 2006).(b) Same as Figure 3a, but for the mean monthly day‐nighttemperature differences for forest and nonforest regionsof the study area. The PRODES forest cover map (years2000–2005) was used to separate forest and nonforest areas,using a threshold of 0.5 fractional cover. (c) Monthly meanprecipitation (mm/month) during 2000–2005 from theTRMM 3B43 product [Kummerow et al., 1998], and monthlyrelative humidity at 1000 hPa (in %) from the AIRS Level 3 V5(AIRX3STM) Retrieval product. A mean for the years 2003–2005 was used.


6 of 16

Figure 4. (left) Maps of (a) daytime LST (K), (c) nighttime LST, and (e) day‐night LST difference inAugust (month of largest day‐night LST difference), averaged during 2000–2005, and (right) their scat-terplots versus forest cover from PRODES averaged over 2000–2005. Also shown in Figure 4 (right) arelinear regression fits (solid line) of (b) daytime LST, (d) nighttime LST, and (f) day‐night difference as afunction of PRODES forest cover. The regression statistics were: (1) for daytime LST, a slope of −6.8,intercept of 306.8, and p < 0.01; (2) for nighttime LST, a slope of 2.5, intercept of 290.9, and p < 0.01;(3) for the day‐night LST difference, a slope of −9.3, intercept of 15.9, and p < 0.01. In total, 30436 pixelswere used for the above regressions.


7 of 16

transpiration even as precipitation levels began to increase.The higher humidity levels for forest as compared withnonforest observed during September and October maydecrease longwave cooling at night (as described above) andthus contribute to the smaller day‐night LST differencesobserved during these months and during the following wetseason.[30] The correlation coefficients between PRODES forest

cover and day LST, night LST, and day‐night LST differ-ence for the PRODES forest domain are given in Table 1.For each pixel we used the monthly means over 2000–2005for LSTs and the annual means over 2000–2005 for thePRODES forest cover product. Pixels with missing or poorquality LST data for all years (2000–2005) were excludedfrom the comparison with PRODES, based on the qualityflag in LST product MOD11B1. The highest correlations forday LST and day‐night LST difference occurred at the endof the dry season (in August), while the highest correlationfor night LST occurred a month earlier (July). Figures 4b,4d, and 4f show the relationship between day LST, nightLST, and day‐night LST difference with the PRODES forestcover in August. For both the day LST and the day‐night LSTdifference, a negative correlation of −0.83 was observed.Night LST had a positive correlation of 0.59 with forestcover. The relationship between land cover and LST wasweaker during the wet season, with correlations in January,for example, of −0.69 for day LST, 0.14 for night LST, and−0.66 for the day‐night LST difference.

3.2. LST and Forest Cover Change

[31] Moderate losses of forest cover were widely distrib-uted in the north central part of Mato Grosso, to the west ofXingu Indigenous Reserve (Figures 2b and 2c). Figure 5shows that some of the largest increases in day LST andday‐night LST difference occurred in these areas. The mapof change in night LST (Figure 5c) shows a more homo-geneous pattern, with increases distributed widely in thesouthern part of the domain.

[32] Correlation coefficients between the change in forestcover during this period (forest fraction/yr) and the changein day LST, night LST, and day‐night LST difference (K/yr)are shown in Table 1, for both the PRODES and Hansenet al. [2008] deforestation products. The LST slope was notcalculated for cloud contaminated pixels, because the influ-ence of these errors on the slope was large in our data set ofonly six years (2000–2005). During the 3 month period fromJanuary–March, 25437 out of 57600 pixels in our study area(44.2%) were contaminated. Contamination during the dry sea-son, from July to September, was considerably lower (0.1%).[33] The seasonal pattern of correlation for forest cover

change was similar to that observed for forest cover: thehighest correlations with the change in day LST and day‐night LST difference occurred at end of the dry season inJuly and August for both the PRODES and Hansen et al.[2008] estimates. The correlation for night LST was high-est during June and July for PRODES and July and Augustfor Hansen et al. [2008], when precipitation was at or nearminimum levels. In Figures 5b, 5d, and 5f the relationshipbetween the change in LST and the change in PRODES forestcover are shown for the month of August. Both the changein day LST and day‐night LST difference were negativelycorrelated with the changes in forest cover, with correlationcoefficients of −0.59 and −0.65, respectively. While the day‐night LST difference performed better than the day LSTfor predicting patterns of deforestation from PRODES, thereverse occurred for the Hansen et al. [2008] product, withcorrelations of −0.70 for day and −0.67 for the day‐nightLST difference. During other dry season months, however,including July and September, correlations with the day‐night LST difference exceeded correlations with day LST forthe Hansen et al. [2008] product (Table 1).

3.3. Feature Selection

[34] Correlations between the change in day and day‐night LST difference and the change in forest cover werestrongest during the dry season, and especially during the

Table 1. Monthly Relationships Between Physical Properties, Land Surface Temperature, and Land Surface Temperature Change

Month

Physical PropertiesCorrelation With PRODES

Forest CoveraCorrelation With PRODESForest Cover Changeb

CorrelationWithHansen et al.[2008] Forest Cover Changec

Precipitationd

(mm/month)

DayLST(K)

NightLST(K)

Day‐NightLST(K)

DayLST

NightLST

Day‐NightLST

DayLST

NightLST

Day‐NightLST

DayLST

NightLST

Day‐NightLST

Jan 291.7 299.1 293.9 5.2 −0.69 0.14 −0.66 −0.12 0.05 −0.15 −0.23 0.19 −0.20Feb 269.1 299.0 294.7 4.3 −0.73 −0.05 −0.50 −0.11 −0.01 −0.01 0.06 −0.09 0.08Mar 263.1 298.8 294.1 4.7 −0.71 −0.09 −0.58 −0.09 0.03 −0.02 −0.03 0.06 0.02Apr 125.8 299.8 293.8 6.0 −0.75 0.10 −0.72 −0.14 0.05 −0.21 0.02 0.17 −0.11May 58.8 299.4 293.1 6.3 −0.71 0.41 −0.75 −0.37 0.02 −0.22 −0.31 −0.08 −0.18Jun 16.0 300.7 291.9 8.8 −0.78 0.57 −0.77 −0.45 0.19 −0.50 −0.35 0.16 −0.39Jul 16.5 301.2 291.1 10.1 −0.75 0.60 −0.80 −0.60 0.25 −0.63 −0.53 0.24 −0.56Aug 24.1 303.6 291.6 12.0 −0.83 0.59 −0.83 −0.59 0.10 −0.65 −0.70 0.23 −0.67Sep 70.5 304.1 293.5 10.6 −0.82 0.47 −0.83 −0.52 0.02 −0.51 −0.52 0.16 −0.56Oct 155.4 303.0 294.1 8.9 −0.79 0.13 −0.80 −0.09 −0.02 −0.08 0.07 0.13 −0.08Nov 189.2 301.8 294.1 7.7 −0.78 0.12 −0.78 −0.28 −0.11 −0.21 −0.32 −0.18 −0.19Dec 279.7 300.6 294.0 6.6 −0.74 0.10 −0.72 −0.20 0.06 −0.06 −0.29 0.10 −0.29aSpatial correlation of LST with the PRODES forest cover product. For both products we used the mean of the years 2000–2005. Pixels with errors in land

surface temperatures due to cloud contamination for all years (2000–2005) were excluded from the comparison with the PRODES forest cover product.bSpatial correlation of the slope of LST with the slope of PRODES forest cover for the years 2000–2005. LST slopes were not calculated for cloud‐

contaminated pixels for the years 2000–2005. The influence of these errors on the slope is large because the data set covers only 6 years.cSpatial correlation of the slope of LST with the Hansen et al. [2008] forest cover loss for the years 2000–2005.dThe 3B43 TRRM monthly accumulated surface rainfall (mm/month).


8 of 16

Figure 5. (left) Spatial distribution of the August (month of largest day‐night LST difference) LST slopein unit K/yr from 2000 to 2005, for (a) daytime LST, (c) nighttime LST, and (e) day‐night LST difference,and (right) the corresponding scatterplots against forest cover change from PRODES during 2000–2005.The results for linear regression fit (solid line) were: (b) for daytime LST, slope of −7.34, intercept of0.37, and p < 0.01; (d) for nighttime LST, slope of 0.46, intercept of 0.20, and p < 0.01; (f) for day‐nightLST difference, slope of −7.80, intercept of 0.17, and p < 0.01. In total, 30436 pixels were used for theabove regressions.


9 of 16

months of July, August, and September in Mato Grosso,as described above (section 3.2 and Table 1). To developan approach that could be applied more widely across thetropics, we needed to develop an algorithm that could locallyidentify the seasonal period of the greatest drought stress. Tofind this period, we identified the period of maximum day‐night LST difference in each grid cell, since Table 1 showsthat during the months with the highest correlations betweenLST and PRODES and Hansen et al. [2008] forest coverchange, the difference between day and night LST is thelargest. First, we used all available observations from our2000–2005 time series to construct a monthly mean annualcycle of day‐night LST difference at each pixel. In a secondstep we found the monthly maximum day‐night LST differ-ence from this climatology, and its corresponding month.This month was defined as the center of a 3 month window

that we then used to sample and calculate the 3 monthaverages of satellite products for our logistic regressionmodel, including day LST, night LST, day‐night LST dif-ference, and NDVI.[35] Using the above described feature selection method,

the correlation coefficients showed a small improvementover the fixed monthly correlation coefficients shown inTable 1. For the correlation with PRODES forest cover, ouralgorithm showed correlation coefficients of −0.83 for dayLST, 0.56 for night LST and −0.84 for day‐night LST dif-ference. The correlation coefficients with forest cover changewere −0.67 and −0.71 (day LST), 0.12 and 0.25 (night LST)and −0.68 and −0.70 (day‐night LST difference) for PRODESand Hansen et al. [2008] products, respectively. Figure 6ashows the month of maximum day‐night LST difference forthe study area, the state of Mato Grosso. For most pixels, this

Figure 6. (a) The month corresponding to the maximum day and night land surface temperature differ-ence for the study area. (b) The maximum monthly day‐night land surface temperature difference (K)averaged over the years 2000–2005.

Table 2. Logistic Regression Model Performance With Different Sets of Satellite Input Data

Experiment

Full Domaina Active Deforestation Domainb

Correlation (r) Area Predicted/Observed Correlation (r) Area Predicted/Observed

LST Sampled During the Dry SeasonDay 0.58 1.51 0.58 2.48Night 0.11 0.82 0.19 −1.09Day‐night 0.64 1.21 0.63 1.41

D‐N LST Sampled During Different Seasonal PeriodsAnnual mean 0.48 0.97 0.58 −0.19Wet season 0.05 1.79 0.08 0.40

NDVI Sampled During Different Seasonal PeriodsDry seasonc 0.80 1.23 0.78 1.24Annual mean 0.66 1.41 0.66 0.93Wet seasond −0.03 0.62 0.03 −1.68

NDVI and D‐N LST Sampled During Different Seasonal PeriodsDry season 0.80 1.22 0.78 1.28Annual mean 0.70 1.20 0.71 0.66Wet season 0.02 1.85 0.07 −0.94

aThe full domain corresponds to the area where PRODES data are available.bThe active deforestation domain corresponds to the pixels where PRODES predicted any deforestation.cThe dry season corresponds to the mean of a 3 month window, with the month of maximum day‐night LST difference as center (see section 3.3).dThe wet season corresponds to the 3 month window that is exactly 6 months out of phase with the dry season.


10 of 16

month corresponded to the end of the dry season (August). Inthe northern part of the study area there were some pixelswhere the maximum in day‐night LST difference occurred1 month earlier, whereas in the south many pixels reached amaximum later in the year, during September, October, andNovember. These differences can likely be explained byregional differences in climate. The maximum day‐night LSTis shown in Figure 6b. Large values were found in the southernpart of the study area, whereas in the north and the wetlands ofthe Pantanal lower values were observed.

3.4. Logistic Regression Model Comparisons

[36] The performances of the logistic regression model topredict forest cover change for different experiments, as

described in section 2.3, are shown in Table 2. We useddifferent seasonal periods of satellite observations to con-struct the logistic regression models. The “dry season” cor-responds to sampling the satellite observations according tothe metric we developed in section 3.3. The wet season cor-responded to a 3 month period that is exactly 6 months out ofphase with the dry season window defined above, and theannual mean is simply the mean for all the months in a cal-endar year.[37] For LSTs sampled during the dry season, the logistic

regression model using the day‐night LST difference per-formed better (r = 0.64) than the day LST (r = 0.58) or nightLST (r = 0.11) models over the full domain. The ratio of thearea of the predicted versus actual forest cover losses was

Figure 7. (a) Pixels in our study domain where PRODES showed more than 20% of deforestationbetween the years 2000 and 2005. Gray corresponds to the area outside the PRODES domain. (b) Pixelsin our study domain where the NDVI and day‐night LST difference sampled during the dry season pre-dicted more than 20% deforestation between the years 2000 and 2005. (c) Pixels in our study domainwhere the NDVI and day‐night LST difference sampled during the dry season predicted more than 5%of regrowth. (d) A map of the predicted deforestation and regrowth overlaid on the PRODES deforesta-tion in our study area. Shown are the pixels where LST and NDVI, sampled during the dry season, pre-dicted more than 20% of deforestation (red), pixels with more than 5% of predicted regrowth (green), andpixels where PRODES showed more than 20% deforestation but our algorithm predicted no or less than20% deforestation (gray).


11 of 16

1.21 for the day‐night LST difference, indicating that themodel was too sensitive to changes in the input features.Over the active deforestation part of the domain, the day‐night LST difference model also had the highest correlationwith observed forest cover loss. The ratio of predicted toobserved forest cover loss was 1.41, and better than the ratioof the day LST (2.48) or night LST (−1.09). A ratio of 1is ideal, so the day and day‐night LST difference over-predicted forest cover loss in grid cells that PRODESidentified as changing during 2000–2005. The negative ratioof −1.09 (night LST) corresponded to a prediction of moreregrowth than deforested area.[38] The use of information at the end of the dry season,

when day‐night LSTwas at maximum, substantially improvedmodel predictions as compared with models using annualmean or wet season observations. This was true for modelsusing day‐night LST, NDVI, and both day‐night LST andNDVI (Table 2). Models using annual mean observationsshowed the second best performance, with the correlation of0.48 and a ratio of the predicted forest cover change overPRODES change of 0.97 in the whole domain. Models usingwet season observations were substantially degraded.[39] The predictions with NDVI showed a slightly better

correlation than that with the day‐night LST differenceduring the dry season, and the combination of the two setsof observations gave the same correlations (r = 0.80 forthe full domain, and r = 0.78 for the active deforestationdomain). Overall, the differences between the performancefor the full domain and for only the active deforestationdomain were small.

[40] The logistic regression model sampling the observa-tions according to our algorithm (section 3.3) performedbetter than a model using observations from a fixed periodduring the dry season: August, the month of highest cor-relations between LST and PRODES forest cover (Table 1).The prediction using only August observations had a corre-lation coefficient with observed forest cover changes of 0.55and 0.57 for day‐night LST difference, 0.75 and 0.76 forNDVI, and 0.77 and 0.78 for the combination of NDVI andday‐night LST difference (for the full and active deforestationdomains, respectively).[41] For predictions with the combined NDVI and day‐

night LST difference during the dry season, the spatial patternof larger clearing areas (greater than 20% during 2000–2005)agreed reasonable well with PRODES (Figures 7a and 7b),although the logistic regression model predicted that the defor-estation rate was somewhat higher than PRODES estimatesin central part of Mato Grosso (9°S–13°S, 54°W–59°W).[42] Besides detecting deforestation, we also used our

algorithm to explore areas of savanna/forest regrowth. Becauseregrowth occurs over longer timescales than deforestation,we set the detection threshold in Figure 7 to a greater than 5%increase in fraction tree cover (compared to a greater than20% decrease for deforestation). Areas where regrowth mightoccurred according to our algorithm were mostly confinedto the savanna woodlands in the southwestern part of thePRODES domain (Figure 7c).[43] Over our study period, climatic conditions (tempera-

ture and precipitation) that could also potentially explainincreases in plant productivity, showed no trend. The areas ofregrowth shown in Figure 7may have been deforested prior toour study period. In past work, regrowth has sometimesbeen excluded from remote sensing studies using fractionaltree cover to detect rates of deforestation [e.g., Achard et al.,2002; Hansen et al., 2008]. Our analysis suggests thatchanges in woodlands and savannas may also be contributingto global land use change carbon fluxes, although more workis needed with higher‐resolution imagery and field mea-surements to verify these results and disentangle natural andanthropogenic causes of variability and trends in savannastructure and productivity.

3.5. Applicability of LST Metrics to Other Regions andTime Periods

[44] To check if our LST metric was applicable for areasother than the state of Mato Grosso, we compared cumu-lative LST changes with estimates from PRODES (available at

Figure 8. Comparison of cumulative day‐night LSTchanges (in K/yr) with estimates of deforestation fromPRODES (in km2/yr) for states in the Brazilian Amazon dur-ing 2001–2005. The day‐night LST changes were estimatedas the slope of the day‐night LST difference for each CMGgrid cell sampled using our LST metric. Pixels with positiveslopes and correlation coefficients larger than 0.7 (r) weresummed within each state. The linear fit with PRODESdata was described by a slope of 0.40, r equal to 0.98, andn equal to 9. Abbreviations are AC, Acre; AM, Amazonas;AP, Amapá; MA, Maranhão; MT, Mato Grosso; PA, Pará;RO, Rondônia; RR, Roraima; TO, Tocantins. Note that bothx and y axis are plotted on a log scale.

Table 3. Relationships Between Changes in Day‐Night LST andPRODES or Hansen et al. [2008] Deforestation Rates

Region Dry Seasona Annual Mean Wet Seasonb

Legal Amazonc 0.98 0.88 0.37Southeast Asiad 0.67 0.46 0.11

aThe dry season corresponded to the mean of the 3 month window,centered on month of maximum day‐night LST difference (see section 3.3).

bThe wet season corresponded to the 3 month window that was exactly6 months out of phase with the dry season.

cAggregated values from nine Brazilian states were used to compute thecorrelation coefficient: Acre, Amazonas, Amapá, Maranhão, Mato Grosso,Pará, Rondônia, Roraima, and Tocantins.

dSix areas were used: Cambodia, Kalimantan, Malaysia, Sulawesi,Sumatra, and Thailand.


12 of 16

the INPE website: http://www.obt.inpe.br/prodes/) for all ofthe Brazilian states within the Amazon (Figure 8). CumulativeLST changes were estimated as the slope of the day‐nightLST difference for the years 2000–2005 sampled using ourgrid cell specific approach for identifying the dry season.For each state, we took the sum of all pixels within eachspatial domain that had a positive slope and a correlationcoefficient larger than 0.7.[45] In Table 3 we report the statistics for our day‐night

LST metric and for day‐night LST changes derived usingeither annual mean or the 6 month out of phase observations.We also show the results for different countries in SoutheastAsia, based on a comparison with the global deforestationproduct from Hansen et al. [2008] for the years 2000–2005.Although the spatial correlations of the regression werenot as good as for the states across the Brazilian Amazon,a clear improvement occurred when day‐night LST wassampled during the local dry season (i.e., during the sea-sonal period when the day‐night LST differences were mostpronounced).[46] In Figure 9 we compare the PRODES deforestation

rate in km2/yr with our day‐night LSTmetric described abovefor two periods: 2001–2005 and 2006–2009. The day‐nightLST changes decreased during 2006–2009 relative to 2001–2005 for many states within the Amazon, providing anindependent confirmation of lower deforestation levels dur-ing the latter part of this decade as reported by PRODES.For the states which are important from a deforestation per-

spective (Mato Grosso, Pará, Rondônia), we observed largedecreases in the day‐night LST metric for the 2006–2009period.

4. Discussion

[47] Our analysis suggests that information from thermalbands can be useful in monitoring deforestation; especiallywhen sampled during the dry season and when day andnight LST observations are combined. Logistic regressionmodels of forest cover change using LST sampled in thisway were slightly less effective than models using NDVI,but they performed well enough to be useful when NDVIdata for a region is unavailable or noisy. LST observationsmay be particularly useful with satellite spectrometers thatdo not have high‐resolution spectral bands that allow forestimation of NDVI (e.g., Geostationary Operational Envi-ronmental Satellite (GOES)), or in areas where aerosols limitthe usefulness of visible and near‐infrared surface reflec-tance observations.[48] The information content of LST data varied consid-

erably from season to season. High correlations betweenday‐night LST difference data and PRODES and Hansenet al. [2008] forest cover change were observed at the endof the dry season (August in our study area and approxi-mately 1–2 months after minimum monthly precipitation).In contrast, during the wet season the correlationswere significantly lower. The higher amount of cloud‐

Figure 9. Time series of PRODES deforestation rate in km2/yr (gray bars, y axis) and the cumulativechange in day‐night LST (in K/yr) within each region (solid black line, secondary y axis) for nine Brazilianstates (years 2001–2005 and 2006–2009).


13 of 16

contaminated pixels in the wet months like December,January, and February might have played a role here. Evenduring periods without cloud cover, increases in columnwater vapor content during the wet season likely contributedto smaller differences in nighttime cooling between forestand nonforest areas (e.g., Figure 3). Higher correlations werefound in the months when cloud contamination and pre-cipitation were low, surface humidity was at a minimum,and the difference between day and night LSTwas the largest.We found large variations in the correlations between LSTdata and forest cover change during the year. Therefore, itis important that the use of LST data in deforestation clas-sifiers take advantage of seasonal periods with the highestdescriptive power. To predict these months, we developeda simple temporal sampling algorithm based on the largestday‐night LST difference. Different tests showed that databased on the LST during a 3 month window centered on themonth derived from this metric yielded the highest correla-tions with independent data.[49] Besides the importance of the temporal sampling of

LST data, our analysis showed that during the end of the dryseason the day‐night LST difference performed better thanthe day LST for comparisons we made with PRODES;several experiments in our logistic regression model con-firmed this. On the other hand, a comparison with Hansenet al. [2008] forest cover loss data showed slightly highercorrelations for the day LST than the day‐night differenceduring August (but the reverse in July and September). Thereasons for the different performance of day‐night LST dif-ferences with PRODES and Hansen et al. [2008] deforesta-tion products remain unclear and will require further analysiswith higher spatial resolution LST data. Day LST has alreadybeen used in different deforestation classifiers; the approachof using the day‐night LST difference data in deforestationclassifiers is novel and may reduce the sensitivity of LST‐based approaches to interannual variability in climate andto variations in topography. Effects of climate change andtopography are expected to influence both day and nightLSTs in the same direction, so day‐night LST differencesmay be less sensitive to these factors. This reduced sensitivitymay be of particular use in constructing long time series ofland cover change that extend across multiple satellites,including those that have different sensor characteristics.[50] A disadvantage of using the day‐night LST differ-

ence as compared to day LST is that the day‐night LSTdifference may be more sensitive to cloud contaminationbecause two, rather than one, clear‐sky images are needed toconstruct this metric and because cloud cover may be higherin tropical regions at night [Durieux et al., 2003]. During thedry season, and especially the months of July, August, andSeptember, day and night LST products from MODIS hadmore cloud‐free observations than during other periods. Asdescribed above, lower water content in the atmosphericcolumn also probably allows for greater nighttime coolingand thus the potential for greater discrimination betweenhigh and low stature vegetation. Thus, sampling LST duringthe end of the dry season using the algorithm we developedmay have advantages considering both surface biophysicaland atmospheric contamination perspectives. In future work,by combining Aqua and Terra observations (with up to fouroverpasses per day) it may be possible to partly resolvesome of issues related to cloud contamination.

[51] The region of focus for this study was in the arc ofdeforestation in the state of Mato Grosso, Brazil. Defores-tation and climatic characteristics in this region are wellsuited for MODIS‐based deforestation monitoring, becauselong dry seasons and low‐stature transition forest typesenable the mechanized clearing of forest cover [Morton et al.,2005]. These large clearings are more easily detected withmoderate resolution remote sensing products, and cloudcover is much lower during the dry season, increasing thechance for cloud‐free MODIS imagery. In other importanthumid tropical regions in Africa and Asia the dry season isconsiderably weaker, limiting cloud-free imagery [Hansenet al., 2009], and also the sizes of tropical forest clearingsmay be smaller than those that typically occur in MatoGrosso, Brazil. Large mechanized clearings are used for thecreation of pastures and croplands in Mato Grosso [Mortonet al., 2006], whereas clearings in Asia and especially Africaare smaller in size and associated with logging activities andsmaller‐scale agricultural complexes.[52] Although some results of our LST metric in other

areas (Brazilian Amazon and Southeast Asia) were promising(Figures 8 and 9), more work with higher‐resolution LST datais needed to assess whether our approach will be useful inregions that have smaller clearing sizes. In principal, subpixelchanges in forest cover should impart changes in LST that aredetectable in moderate resolution observations; the linearityof these responses are not well understood and require furtherstudy.[53] Use of the coarse resolution climate modeling grid

(CMG) products here was a useful first step for evaluatingthe information content of different LST measurements as afunction of overpass time and monthly sampling interval. Animportant next step toward the development of an operationalpantropical forest classifier is to assess the performance ofthese different types of observations with 1 km MODIS LSTdata, 90 m Advanced Spaceborne Thermal Emission andReflection Radiometer (ASTER) observations, and in future,with 60 m observations from the planned HyperspectralInfrared Imager (HyspIRI) Mission. Higher‐resolution LSTdata also are likely to be more effective in identifying theexact geographic location of deforestation. MODIS LST dataare appropriate for rapid identification of the location ofdeforestation areas and trends in deforestation dynamics, butcannot be seen as a replacement for higher‐resolution systemsbased on Landsat and other higher‐resolution surfacereflectance imagery.

5. Conclusions

[54] LST data can be a useful information source inclassification models for distinguishing between forest andnonforest areas and for quantifying deforestation in a con-tinuous way. The information content of the day LST andday‐night LST difference observations varied considerablyover an annual cycle. At the end of the dry season, whenprecipitation was low, surface humidity was at a minimum,and the difference between day and night LST was large, thehighest correlations were found with the PRODES andHansen et al. [2008] forest cover change in the Brazilianstate of Mato Grosso. Based on this analysis, we developeda metric to monitor forest cover change in the humid tropics.Besides the importance of sampling LST data during the


14 of 16

months with the highest descriptive power, we found thatthe day‐night LST difference in general improved the abilityof classification models to distinguish forest and nonforestareas as compared with models that only use day LST. Wewere able to show using this approach that changes in day‐night LSTs were highly correlated with the spatial patternof deforestation across different Brazilian states within theAmazon. Changes in day‐night LSTs were lower within theAmazon during 2006–2009 relative to 2001–2005, provid-ing independent confirmation of decreases in the rate ofdeforestation during the latter part of this decade as reportedby PRODES.[55] Further research is needed to optimize our day‐night

LST difference algorithm in detecting tropical forest coverchange; the use of higher spatial and temporal resolutionLST data needs to be explored, because the actual clearingsize of forest in often smaller than the pixel size of the coarseresolution (0.05° × 0.05°) LST data used in this study. Longertime series of LST from MODIS, that now extend overa decade in length, may increase our ability to use LSTobservations in effective ways to detect long‐term changes inforest cover. Next steps are to validate this approach usinghigher‐resolution Landsat observations and to extend thisapproach globally.

[56] Acknowledgments. This work was supported by NASA grantsNNX08AF64G, NNX08AG13G, and NNX08AR69G, and NSF grantIIS‐0431085. We obtained the MODIS observations from the LandProcesses Distributed Active Archive Center (LP DAAC) located at theU.S. Geological Survey (USGS) Earth Resources Observation and Science(EROS) Center (https://lpdaac.usgs.gov).

ReferencesAchard, F., et al. (2002), Determination of deforestation rates of the world’shumid tropical forests, Science, 297(5583), 999–1002, doi:10.1126/science.1070656.

Achard, F., et al. (2007), Pan‐tropical monitoring of deforestation, Environ.Res. Lett., 2, 045022, doi:10.1088/1748-9326/2/4/045022.

Anderson, L. O., et al. (2005), Assessment of deforestation in near real timeover the Brazilian Amazon using multitemporal fraction images derivedfrom terra MODIS, IEEE Geosci. Remote Sens. Lett., 2(3), 315–318,doi:10.1109/LGRS.2005.850364.

Andreux, F. G., and C. C. Cerri (1989), Current trends in the research onsoil changes due to deforestation, burning and cultivation in the Braziliantropics, Toxicol. Environ. Chem., 20(1), 275–283, doi:10.1080/02772248909357387.

Aumann, H. H., et al. (2003), AIRS/AMSU/HSB on the aqua mission:Design, science objectives, data products, and processing systems, IEEETrans. Geosci. Remote Sens. , 41(2), 253–264, doi:10.1109/TGRS.2002.808356.

Bala, G., et al. (2007), Combined climate and carbon‐cycle effects of large‐scale deforestation, Proc. Natl. Acad. Sci. U. S. A., 104(16), 6550–6555,doi:10.1073/pnas.0608998104.

Borak, J. S., E. F. Lambin, and A. H. Strahler (2000), The use of temporalmetrics for land cover change detection at coarse spatial scales, Int. J.Remote Sens., 21, 1415–1432, doi:10.1080/014311600210245.

Bruno, R. D., et al. (2006), Soil moisture dynamics in an eastern Amazo-nian tropical forest, Hydrol. Process., 20(12), 2477–2489, doi:10.1002/hyp.6211.

Canadell, J. G., et al. (2007), Contributions to accelerating atmosphericCO2 growth from economic activity, carbon intensity, and efficiency ofnatural sinks, Proc. Natl. Acad. Sci. U. S. A., 104(47), 18,866–18,870,doi:10.1073/pnas.0702737104.

Carlson, T., and R. Gillies (1994), A method to make use of thermal infra-red temperature and NDVI measurements to infer surface soil water con-tent and fractional vegetation cover, Remote Sens. Rev., 9, 161–173.

Carreiras, J. M. B., et al. (2002), Fraction images derived from SPOT‐4VEGETATION data to assess land‐cover change over the state of MatoGrosso, Brazil, Int. J. Remote Sens., 23(23), 4979–4983, doi:10.1080/0143116021000016743.

Collatz, G. J., et al. (2000), A mechanism for the influence of vegetation onthe response of the diurnal temperature range to changing climate,Geophys. Res. Lett., 27(20), 3381–3384, doi:10.1029/1999GL010947.

Davidson, E. A., et al. (2007), Recuperation of nitrogen cycling in Amazonianforests following agricultural abandonment, Nature, 447, 995–998,doi:10.1038/nature05900.

DeFries, R. S., and J. R. G. Townshend (1994), NDVI‐derived land‐coverclassifications at a global scale, Int. J. Remote Sens., 15(17), 3567–3586,doi:10.1080/01431169408954345.

DeFries, R. S., et al. (1995), Mapping the land surface for globalatmosphere‐biosphere models: Toward continuous distributions ofvegetation’s functional properties, J. Geophys. Res., 100(D10), 20,867–20,882, doi:10.1029/95JD01536.

DeFries, R. S., et al. (2000), Global continuous fields of vegetation charac-teristics: A linear mixture model applied to multi‐year 8 km AVHRR data,Int. J. Remote Sens., 21(6–7), 1389–1414, doi:10.1080/014311600210236.

DeFries, R. S., et al. (2002), Carbon emissions from tropical deforestationand regrowth based on satellite observations for the 1980s and 1990s,Proc. Natl. Acad. Sci. U. S. A., 99(22), 14,256–14,261, doi:10.1073/pnas.182560099.

DeFries, R. S., et al. (2004), Land‐use choices: Balancing human needs andecosystem function, Front. Ecol. Environ, 2(5), 249–257, doi:10.1890/1540-9295(2004)002[0249:LCBHNA]2.0.CO;2.

DeFries, R., et al. (2007), Earth observations for estimating greenhousegas emissions from deforestation in developing countries, Environ. Sci.Policy, 10, 385–394, doi:10.1016/j.envsci.2007.01.010.

Durieux, L., L. A. T. Machado, and H. Laurent (2003), The impact ofdeforestation on cloud cover over the Amazon arc of deforestation,Remote Sens. Environ., 86, 132–140, doi:10.1016/S0034-4257(03)00095-6.

Eidenshink, J. C., and J. L. Faundeen (1994), The 1 km AVHRR globalland data set—1st stages in implementation, Int. J. Remote Sens., 15(17),3443–3462, doi:10.1080/01431169408954339.

Foley, J. A., et al. (2005), Global consequences of land use, Science,309(5734), 570–574, doi:10.1126/science.1111772.

Food and Agriculture Organization (2006), Global Forest ResourcesAssessment 2005, FAO For. Pap. 147, U. N., Rome.

Friedl, M. A. (2002), Forward and inverse modeling of land surface energybalance using surface temperature measurements, Remote Sens. Environ.,79(2–3), 344–354, doi:10.1016/S0034-4257(01)00284-X.

Friedl, M. A., and C. E. Brodley (1997), Decision tree classification ofland cover from remotely sensed data, Remote Sens. Environ., 61(3),399–409, doi:10.1016/S0034-4257(97)00049-7.

Friedl, M. A., et al. (2002), Global land cover mapping from MODIS:Algorithms and early results, Remote Sens. Environ., 83(1–2), 287–302,doi:10.1016/S0034-4257(02)00078-0.

Goetz, S. J., et al. (2009), Mapping and monitoring carbon stocks with sat-ellite observations: A comparison of methods, Carbon Balance Manag.,4(2), doi:10.1186/1750-0680-4-2.

Gopal, S., and C. Woodcock (1996), Remote sensing of forest change usingartificial neural networks, IEEE Trans. Geosci. Remote Sens., 34(2),398–404, doi:10.1109/36.485117.

Goulden, M. L., et al. (2006), Nocturnal cold air drainage and pooling ina tropical forest, J. Geophys. Res., 111, D08S04, doi:10.1029/2005JD006037.

Hamilton, L. S., and P. N. King (1983), Tropical Forested Watersheds:Hydrologic and Soils Response to Major Uses or Conversions, pp. 137–168, Westview, Boulder, Colo.

Hansen, M. C., and R. S. DeFries (2004), Detecting long‐term global forestchange using continuous fields of tree‐cover maps from 8‐km advancedvery high resolution radiometer (AVHRR) data for the years 1982–99,Ecosystems, 7(7), 695–716, doi:10.1007/s10021-004-0243-3.

Hansen, M. C., et al. (2000), Global land cover classification at 1km spa-tial resolution using a classification tree approach, Int. J. Remote Sens.,21(6–7), 1331–1364, doi:10.1080/014311600210209.

Hansen, M. C., et al. (2008), Humid tropical forest clearing from 2000 to2005 quantified by using multitemporal and multiresolution remotelysensed data, Proc. Natl. Acad. Sci. U. S. A., 105(27), 9439–9444,doi:10.1073/pnas.0804042105.

Hansen, M. C., et al. (2009), Quantifying changes in the rates of forestclearing in Indonesia from 1990 to 2005 using remotely sensed data sets,Environ. Res. Lett., 4(3), 034001, doi:10.1088/1748-9326/4/3/034001.

Holben, B. (1986), Characteristics of maximum‐value composite imagesfrom temporal AVHRR data, Int. J. Remote Sens., 7, 1417–1434,doi:10.1080/01431168608948945.

Hosmer, D. W., and S. Lemeshow (2004), Applied Logistic Regression,Wiley, New York.

Houghton, R. A. (1991), Tropical deforestation and atmospheric carbon‐dioxide, Clim. Change, 19(1–2), 99–118, doi:10.1007/BF00142217.


15 of 16

Houghton, R. A., et al. (2000), Annual fluxes or carbon from deforestationand regrowth in the Brazilian Amazon, Nature, 403(6767), 301–304,doi:10.1038/35002062.

Huete, A. R., et al. (2002), Overview of the radiometric and biophysicalperformance of the vegetation indices, Remote Sens. Environ., 83(1–2),195–213, doi:10.1016/S0034-4257(02)00096-2.

Instituto Nacional de Pesquisas Espaciais (INPE) (2002), Deforestationestimates in the BrazilianAmazon, São José dos Campos,Minist. da Ciênc.e Tecnol., Brasilia. (Available at <http://www.obt.inpe.br/prodes/)>

Jasinski, E., et al. (2005), Physical landscape correlates of the expansion ofmechanized agriculture in Mato Grosso, Brazil, Earth Interact., 9, 1–18,doi:10.1175/EI143.1.

Kummerow, C., et al. (1998), The Tropical Rainfall Measuring Mission(TRMM) sensor package, J. Atmos. Oceanic Technol., 15(3), 809–817,doi:10.1175/1520-0426(1998)015<0809:TTRMMT>2.0.CO;2.

Lambin, E. F., and D. Ehrlich (1995), The vegetation indices and surfacetemperature for land‐cover mapping at broad spatial scales, Int. J. RemoteSens., 16, 573–579, doi:10.1080/01431169508954423.

Los, S. O., et al. (1994), A global 1‐degree‐by‐1‐degree NDVI data set forclimate studies derived from the GIMMS continental NDVI data, Int. J.Remote Sens., 15(17), 3493–3518, doi:10.1080/01431169408954342.

Mildrexler, D. J., et al. (2007), A new satellite‐based methodology for con-tinental‐scale disturbance detection, Ecol. Appl., 17(1), 235–250,doi:10.1890/1051-0761(2007)017[0235:ANSMFC]2.0.CO;2.

Morton, D. C., et al. (2005), Rapid assessment of annual deforestation inthe Brazilian Amazon using MODIS data, Earth Interact., 9, 1–22,doi:10.1175/EI139.1.

Morton, D. C., et al. (2006), Cropland expansion changes deforestationdynamics in the southern Brazilian Amazon, Proc. Natl. Acad. Sci. U. S. A.,103(39), 14,637–14,641, doi:10.1073/pnas.0606377103.

Nabuurs, G. J., et al. (2007), Forestry, in Climate Change 2007: Mitigation.Contribution of Working Group III to the Fourth Assessment Report ofthe Intergovernmental Panel on Climate Change, edited by B. Metz et al.,pp. 547–561, Cambridge Univ. Press, Cambridge, U. K.

Nelder, J. A., and R. W. M. Wedderburn (1972), Generalized linear models,J. R. Stat. Soc. A, 135, 370–384.

Nemani, R. R., and S. W. Running (1997), Land cover characterizationusing multitemporal red, near‐IR, and thermal‐IR data from NOAA/AVHRR, Ecol. Appl., 7, 79–90, doi:10.1890/1051-0761(1997)007[0079:LCCUMR]2.0.CO;2.

Nemani, R., et al. (1993), Developing satellite‐derived estimates of surfacemoisture status, J. Appl. Meteorol., 32(3), 548–557, doi:10.1175/1520-0450(1993)032<0548:DSDEOS>2.0.CO;2.

Nobre, C. A., et al. (1991), Amazonian deforestation and regional climatechange, J. Clim., 4(10), 957–988, doi:10.1175/1520-0442(1991)004<0957:ADARCC>2.0.CO;2.

Pongratz, J., et al. (2006), The impact of land cover change on surfaceenergy and water balance in Mato Grosso, Brazil, Earth Interact., 10,1–17, doi:10.1175/EI176.1.

Price, J. C. (1984), Land surface temperature measurements from the splitwindow channels of the NOAA 7 advanced very high resolution radi-ometer, J. Geophys. Res. , 89(D5), 7231–7237, doi:10.1029/JD089iD05p07231.

Roy, D. P., P. Kennedy, and S. Folving (1997), Combination of the Nor-malized Difference Vegetation Index and surface temperature forregional scale European Forest cover mapping using AVHRR data, Int.J. Remote Sens., 18, 1189–1195, doi:10.1080/014311697218665.

Sala, O. E., et al. (2000), Biodiversity: Global biodiversity scenarios for theyear 2100, Science, 287(5459), 1770–1774, doi:10.1126/science.287.5459.1770.

Salomonson, V. V., et al. (1989), MODIS: Advanced facility instrument forstudies of the Earth as a system, IEEE Trans. Geosci. Remote Sens., 27(2),145–153, doi:10.1109/36.20292.

Shimabukuro, Y. E., et al. (1998), Vegetation index and spectral linear mix-ing model for monitoring the Pantanal region, Pesquisa Agropecu. Bras.,33, 1729–1737.

Shukla, J., et al. (1990), Amazon deforestation and climate change, Science,247(4948), 1322–1325, doi:10.1126/science.247.4948.1322.

Smith, R. C. G., and B. J. Choudhury (1991), Analyses of normalized dif-ference and surface temperature observations over southeastern Australia,Int. J. Remote Sens., 12, 2021–2044, doi:10.1080/01431169108955234.

United Nations Framework Convention for Climate Change (2005), EleventhConference of Parties (COP), Agenda item 6: Reducing emissions fromdeforestation in developing countries, Geneva, Switzerland.

Wan, Z., Y. Zhang, Q. Zhang, and Z. L. Li (2002), Validation of the land‐surface temperature products retrieved from Terra Moderate ResolutionImaging Spectroradiometer data, Remote Sens. Environ., 83, 163–180,doi:10.1016/S0034-4257(02)00093-7.

A. J. Frank and P. Smyth, Department of Computer Science, Universityof California, Donald Bren Hall, Irvine, CA 92697, USA. ([email protected]; [email protected])M. L. Goulden, Y. Jin, and J. T. Randerson, Department of Earth System

Science, University of California, Croul Hall, Irvine, CA 92697, USA.([email protected]; [email protected]; [email protected])G. R. van der Werf and T. T. van Leeuwen, Department of Hydrology

and Geo‐environmental Sciences, VU University Amsterdam, DeBoelelaan 1085, NL‐1081 HV Amsterdam, Netherlands. ([email protected]; [email protected])


16 of 16

optimal use of land surface temperature data to detect ... · optimal use of land surface...

Documents